Ultralytics v8.3.176 Release: Major Experiment Tracking, Deployment, and Internals Upgrades!
Summary
We’re excited to announce the release of Ultralytics v8.3.176! This version brings enhanced experiment traceability with Comet ML, streamlined YOLOE export and class handling, a core tracking refactor for cleaner code and future flexibility, plus a brand new CoreML tutorial for faster YOLO11 deployment on Apple devices. No breaking changes—just a smoother, more productive workflow for everyone!
Get the details below and see how these updates can help you track, deploy, and iterate faster.
New Features & Improvements
Comet ML Integration (Priority)
- Automatic results logging:
results.csv
andargs.yaml
artifacts are now automatically logged to Comet ML at the end of training for Detect and OBB (oriented bounding box) models, improving reproducibility and sharing. - Enhanced asset/table logging: New helpers for logging assets and tables, plus improved plot logging and code clarity for detection/OBB experiments.
Comet Integration PR by @yaricom
YOLOE Export and Class-Handling
- Export-only fusion: Prompt embeddings for YOLOE are now fused only during export (not during general fuse steps), making model exports more robust.
- Flexible class assignment: You can now set new classes even after fusing, lifting previous restrictions for YOLOE models.
Unified Tracking Internals
- Unified Results object: Both BOT-SORT and BYTETracker now use a unified Results object for initializing and updating tracks, replacing previous per-array logic.
- Cleaner, more maintainable code: Duplicate code is removed, consistency improved, and future tracking enhancements become easier—no changes to typical user APIs.
Results Object for Tracker PR by @Laughing-q
Documentation Update: CoreML Video Tutorial
- Faster Apple deployments: The CoreML guide now includes a YouTube tutorial demonstrating how to export YOLO11 for up to 2x faster inference on iOS/macOS devices—speed up your Apple workflows now!
CoreML Video Docs PR by @RizwanMunawar
Why This Matters
- Better experiment traceability: Automatic metric/config logging to Comet ML enables more transparent, reproducible, and shareable experiments.
- Smoother YOLOE deployment: Reduced fusion-related errors during export and easier class customization streamline advanced YOLOE use cases.
- Cleaner tracking code: A unified batch results approach future-proofs tracker integrations and extensions, while minimizing bugs.
- Accelerated Apple device workflows: New CoreML resources help you deploy YOLO11 on iOS/macOS devices faster than ever.
No expected breaking changes; most users should upgrade smoothly.
What’s Changed (with PR and Author Links)
- Add CoreML video tutorial to docs by @RizwanMunawar
- YOLOE: Scope prompt embedding fuse to export process by @Laughing-q
- Use Results object to filter candidates for tracker by @Laughing-q
- Comet integration improvements for YOLO Detect and OBB models by @yaricom
Full Changelog: Compare v8.3.175…v8.3.176
Full Release Details: Ultralytics v8.3.176 Release Page
How to Upgrade & Get Involved
Update with:
pip install --upgrade ultralytics
We encourage you to try out the new features, especially the Comet ML integration and YOLOE export/class improvements! Watch the new CoreML tutorial to optimize your Apple device deployments.
Your feedback is invaluable—share your thoughts, issues, or ideas in the comments below or on Ultralytics Discussions. This release, like all progress, is thanks to the community and the dedicated Ultralytics team—thank you for being a part of it!